block.h

Go to the documentation of this file.

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */
 
#ifndef AOM_AV1_ENCODER_BLOCK_H_
#define AOM_AV1_ENCODER_BLOCK_H_
 
#include "av1/common/blockd.h"
#include "av1/common/entropymv.h"
#include "av1/common/entropy.h"
#include "av1/common/enums.h"
#include "av1/common/mvref_common.h"
 
#include "av1/encoder/enc_enums.h"
#if !CONFIG_REALTIME_ONLY
#include "av1/encoder/partition_cnn_weights.h"
#endif
 
#include "av1/encoder/hash_motion.h"
 
#ifdef __cplusplus
extern "C" {
#endif
 
#define MIN_TPL_BSIZE_1D 16
#define MAX_TPL_BLK_IN_SB (MAX_SB_SIZE / MIN_TPL_BSIZE_1D)
#define RD_RECORD_BUFFER_LEN 8
 
#define MAX_TX_TYPE_PROB 1024
 
typedef struct {
  BLOCK_SIZE min_partition_size;
  BLOCK_SIZE max_partition_size;
 
 
  int tpl_data_count;
  int64_t tpl_inter_cost[MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB];
  int64_t tpl_intra_cost[MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB];
  int_mv tpl_mv[MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB][INTER_REFS_PER_FRAME];
  int tpl_stride;
} SuperBlockEnc;
 
typedef struct {
  MB_MODE_INFO mbmi;
  RD_STATS rd_cost;
  int64_t rd;
  int rate_y;
  int rate_uv;
  uint8_t color_index_map[MAX_SB_SQUARE];
  THR_MODES mode_index;
} WinnerModeStats;
 
typedef struct macroblock_plane {
  int16_t *src_diff;
  tran_low_t *dqcoeff;
  tran_low_t *qcoeff;
  tran_low_t *coeff;
  uint16_t *eobs;
  uint8_t *txb_entropy_ctx;
  struct buf_2d src;
 
  const int16_t *quant_fp_QTX;
  const int16_t *round_fp_QTX;
  const int16_t *quant_QTX;
  const int16_t *round_QTX;
  const int16_t *quant_shift_QTX;
  const int16_t *zbin_QTX;
  const int16_t *dequant_QTX;
} MACROBLOCK_PLANE;
 
typedef struct {
  int txb_skip_cost[TXB_SKIP_CONTEXTS][2];
  int base_eob_cost[SIG_COEF_CONTEXTS_EOB][3];
  int base_cost[SIG_COEF_CONTEXTS][8];
  int eob_extra_cost[EOB_COEF_CONTEXTS][2];
  int dc_sign_cost[DC_SIGN_CONTEXTS][2];
  int lps_cost[LEVEL_CONTEXTS][COEFF_BASE_RANGE + 1 + COEFF_BASE_RANGE + 1];
} LV_MAP_COEFF_COST;
 
typedef struct {
  int eob_cost[2][11];
} LV_MAP_EOB_COST;
 
typedef struct {
  tran_low_t *tcoeff[MAX_MB_PLANE];
  uint16_t *eobs[MAX_MB_PLANE];
  uint8_t *entropy_ctx[MAX_MB_PLANE];
} CB_COEFF_BUFFER;
 
typedef struct {
  // TODO(angiebird): Reduce the buffer size according to sb_type
  CANDIDATE_MV ref_mv_stack[MODE_CTX_REF_FRAMES][USABLE_REF_MV_STACK_SIZE];
  uint16_t weight[MODE_CTX_REF_FRAMES][USABLE_REF_MV_STACK_SIZE];
  uint8_t ref_mv_count[MODE_CTX_REF_FRAMES];
  int_mv global_mvs[REF_FRAMES];
  int16_t mode_context[MODE_CTX_REF_FRAMES];
} MB_MODE_INFO_EXT;
 
typedef struct {
  CANDIDATE_MV ref_mv_stack[USABLE_REF_MV_STACK_SIZE];
  uint16_t weight[USABLE_REF_MV_STACK_SIZE];
  uint8_t ref_mv_count;
  // TODO(Ravi/Remya): Reduce the buffer size of global_mvs
  int_mv global_mvs[REF_FRAMES];
  int16_t mode_context;
  uint16_t cb_offset[PLANE_TYPES];
} MB_MODE_INFO_EXT_FRAME;
 
typedef struct {
  TX_SIZE tx_size;
  TX_SIZE inter_tx_size[INTER_TX_SIZE_BUF_LEN];
  uint8_t blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE];
  uint8_t tx_type_map[MAX_MIB_SIZE * MAX_MIB_SIZE];
  RD_STATS rd_stats;
  uint32_t hash_value;
} MB_RD_INFO;
 
typedef struct {
  MB_RD_INFO mb_rd_info[RD_RECORD_BUFFER_LEN];
  int index_start;
  int num;
  CRC32C crc_calculator;
} MB_RD_RECORD;
 
#define MAX_COMP_RD_STATS 64
 
typedef struct {
  int32_t rate[COMPOUND_TYPES];
  int64_t dist[COMPOUND_TYPES];
  int32_t model_rate[COMPOUND_TYPES];
  int64_t model_dist[COMPOUND_TYPES];
  int comp_rs2[COMPOUND_TYPES];
  int_mv mv[2];
  MV_REFERENCE_FRAME ref_frames[2];
  PREDICTION_MODE mode;
  int_interpfilters filter;
  int ref_mv_idx;
  int is_global[2];
  INTERINTER_COMPOUND_DATA interinter_comp;
} COMP_RD_STATS;
 
typedef struct {
  int32_t *wsrc;
  int32_t *mask;
  uint8_t *above_pred;
  uint8_t *left_pred;
} OBMCBuffer;
 
typedef struct {
  uint8_t best_palette_color_map[MAX_PALETTE_SQUARE];
  int kmeans_data_buf[2 * MAX_PALETTE_SQUARE];
} PALETTE_BUFFER;
 
typedef struct {
  uint8_t *pred0;
  uint8_t *pred1;
  int16_t *residual1;
  int16_t *diff10;
  uint8_t *tmp_best_mask_buf;
} CompoundTypeRdBuffers;
 
// TODO(chiyotsai@google.com): Consolidate this with SIMPLE_MOTION_DATA_TREE
typedef struct {
#if !CONFIG_REALTIME_ONLY
  // The following 4 parameters are used for cnn-based partitioning on intra
  // frame.
  int quad_tree_idx;
  int cnn_output_valid;
  float cnn_buffer[CNN_OUT_BUF_SIZE];
  float log_q;
#endif
 
  uint8_t variance_low[105];
} PartitionSearchInfo;
 
typedef struct {
  int use_default_intra_tx_type;
 
  int default_inter_tx_type_prob_thresh;
 
  int prune_2d_txfm_mode;
 
  unsigned int coeff_opt_thresholds[2];
  unsigned int tx_domain_dist_threshold;
  TX_SIZE_SEARCH_METHOD tx_size_search_method;
  unsigned int use_transform_domain_distortion;
  unsigned int skip_txfm_level;
 
  TX_MODE tx_mode_search_type;
 
  unsigned int predict_dc_level;
 
  int use_qm_dist_metric;
 
  int mode_eval_type;
} TxfmSearchParams;
 
#define MAX_NUM_8X8_TXBS ((MAX_MIB_SIZE >> 1) * (MAX_MIB_SIZE >> 1))
#define MAX_NUM_16X16_TXBS ((MAX_MIB_SIZE >> 2) * (MAX_MIB_SIZE >> 2))
#define MAX_NUM_32X32_TXBS ((MAX_MIB_SIZE >> 3) * (MAX_MIB_SIZE >> 3))
#define MAX_NUM_64X64_TXBS ((MAX_MIB_SIZE >> 4) * (MAX_MIB_SIZE >> 4))
 
typedef struct {
  int skip_txfm;
 
  uint8_t blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE];
 
  uint8_t tx_type_map_[MAX_MIB_SIZE * MAX_MIB_SIZE];
 
  MB_RD_RECORD *mb_rd_record;
 
  // TODO(chiyotsai@google.com): Move this to a more appropriate location such
  // as ThreadData.
  unsigned int txb_split_count;
#if CONFIG_SPEED_STATS
  unsigned int tx_search_count;
#endif  // CONFIG_SPEED_STATS
} TxfmSearchInfo;
#undef MAX_NUM_8X8_TXBS
#undef MAX_NUM_16X16_TXBS
#undef MAX_NUM_32X32_TXBS
#undef MAX_NUM_64X64_TXBS
 
typedef struct {
 
  int partition_cost[PARTITION_CONTEXTS][EXT_PARTITION_TYPES];
 
  int mbmode_cost[BLOCK_SIZE_GROUPS][INTRA_MODES];
  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
  int intra_uv_mode_cost[CFL_ALLOWED_TYPES][INTRA_MODES][UV_INTRA_MODES];
  int filter_intra_cost[BLOCK_SIZES_ALL][2];
  int filter_intra_mode_cost[FILTER_INTRA_MODES];
  int angle_delta_cost[DIRECTIONAL_MODES][2 * MAX_ANGLE_DELTA + 1];
 
  int cfl_cost[CFL_JOINT_SIGNS][CFL_PRED_PLANES][CFL_ALPHABET_SIZE];
 
  int intrabc_cost[2];
 
  int palette_y_size_cost[PALATTE_BSIZE_CTXS][PALETTE_SIZES];
  int palette_uv_size_cost[PALATTE_BSIZE_CTXS][PALETTE_SIZES];
  int palette_y_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
                          [PALETTE_COLORS];
  int palette_uv_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
                           [PALETTE_COLORS];
  int palette_y_mode_cost[PALATTE_BSIZE_CTXS][PALETTE_Y_MODE_CONTEXTS][2];
  int palette_uv_mode_cost[PALETTE_UV_MODE_CONTEXTS][2];
 
  int skip_mode_cost[SKIP_MODE_CONTEXTS][2];
  int newmv_mode_cost[NEWMV_MODE_CONTEXTS][2];
  int zeromv_mode_cost[GLOBALMV_MODE_CONTEXTS][2];
  int refmv_mode_cost[REFMV_MODE_CONTEXTS][2];
  int drl_mode_cost0[DRL_MODE_CONTEXTS][2];
 
  int single_ref_cost[REF_CONTEXTS][SINGLE_REFS - 1][2];
  int comp_inter_cost[COMP_INTER_CONTEXTS][2];
  int comp_ref_type_cost[COMP_REF_TYPE_CONTEXTS]
                        [CDF_SIZE(COMP_REFERENCE_TYPES)];
  int uni_comp_ref_cost[UNI_COMP_REF_CONTEXTS][UNIDIR_COMP_REFS - 1]
                       [CDF_SIZE(2)];
  int comp_ref_cost[REF_CONTEXTS][FWD_REFS - 1][2];
  int comp_bwdref_cost[REF_CONTEXTS][BWD_REFS - 1][2];
 
  int intra_inter_cost[INTRA_INTER_CONTEXTS][2];
  int inter_compound_mode_cost[INTER_MODE_CONTEXTS][INTER_COMPOUND_MODES];
  int compound_type_cost[BLOCK_SIZES_ALL][MASKED_COMPOUND_TYPES];
  int wedge_idx_cost[BLOCK_SIZES_ALL][16];
  int interintra_cost[BLOCK_SIZE_GROUPS][2];
  int wedge_interintra_cost[BLOCK_SIZES_ALL][2];
  int interintra_mode_cost[BLOCK_SIZE_GROUPS][INTERINTRA_MODES];
 
  int comp_idx_cost[COMP_INDEX_CONTEXTS][2];
  int comp_group_idx_cost[COMP_GROUP_IDX_CONTEXTS][2];
 
  int motion_mode_cost[BLOCK_SIZES_ALL][MOTION_MODES];
  int motion_mode_cost1[BLOCK_SIZES_ALL][2];
  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
 
  int skip_txfm_cost[SKIP_CONTEXTS][2];
  int tx_size_cost[TX_SIZES - 1][TX_SIZE_CONTEXTS][TX_SIZES];
  int txfm_partition_cost[TXFM_PARTITION_CONTEXTS][2];
  int inter_tx_type_costs[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES];
  int intra_tx_type_costs[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
                         [TX_TYPES];
 
  int switchable_restore_cost[RESTORE_SWITCHABLE_TYPES];
  int wiener_restore_cost[2];
  int sgrproj_restore_cost[2];
 
  int tmp_pred_cost[SEG_TEMPORAL_PRED_CTXS][2];
  int spatial_pred_cost[SPATIAL_PREDICTION_PROBS][MAX_SEGMENTS];
} ModeCosts;
 
typedef struct {
 
  int nmv_joint_cost[MV_JOINTS];
 
  int nmv_cost_alloc[2][MV_VALS];
  int nmv_cost_hp_alloc[2][MV_VALS];
  int *nmv_cost[2];
  int *nmv_cost_hp[2];
  int **mv_cost_stack;
} MvCosts;
 
typedef struct {
  int joint_mv[MV_JOINTS];
 
  int dv_costs_alloc[2][MV_VALS];
 
  int *dv_costs[2];
} IntraBCMVCosts;
 
typedef struct {
  LV_MAP_COEFF_COST coeff_costs[TX_SIZES][PLANE_TYPES];
  LV_MAP_EOB_COST eob_costs[7][2];
} CoeffCosts;
 
// 4: NEAREST, NEW, NEAR, GLOBAL
#define SINGLE_REF_MODES ((REF_FRAMES - 1) * 4)
 
struct inter_modes_info;
 
typedef struct {
  int num;
  int pts[16];
  int pts_inref[16];
} WARP_SAMPLE_INFO;
 
typedef enum {
  kZeroSad = 0,
  kLowSad = 1,
  kMedSad = 2,
  kHighSad = 3
} SOURCE_SAD;
 
typedef struct {
  SOURCE_SAD source_sad_nonrd;
  SOURCE_SAD source_sad_rd;
  int lighting_change;
  int low_sumdiff;
} CONTENT_STATE_SB;
 
// Structure to hold pixel level gradient info.
typedef struct {
  uint16_t abs_dx_abs_dy_sum;
  int8_t hist_bin_idx;
  bool is_dx_zero;
} PixelLevelGradientInfo;
 
// Structure to hold the variance and log(1 + variance) for 4x4 sub-blocks.
typedef struct {
  double log_var;
  int var;
} Block4x4VarInfo;
 
#ifndef NDEBUG
typedef struct SetOffsetsLoc {
  int mi_row;
  int mi_col;
  BLOCK_SIZE bsize;
} SetOffsetsLoc;
#endif  // NDEBUG
 
typedef struct macroblock {
 
  struct macroblock_plane plane[MAX_MB_PLANE];
 
  MACROBLOCKD e_mbd;
 
  MB_MODE_INFO_EXT mbmi_ext;
 
  MB_MODE_INFO_EXT_FRAME *mbmi_ext_frame;
 
  FRAME_CONTEXT *row_ctx;
  FRAME_CONTEXT *tile_pb_ctx;
 
  CB_COEFF_BUFFER *cb_coef_buff;
  uint16_t cb_offset[PLANE_TYPES];
 
  OBMCBuffer obmc_buffer;
  PALETTE_BUFFER *palette_buffer;
  CompoundTypeRdBuffers comp_rd_buffer;
  CONV_BUF_TYPE *tmp_conv_dst;
 
  uint8_t *tmp_pred_bufs[2];
 
  int qindex;
 
  int delta_qindex;
 
  int rdmult;
 
  int intra_sb_rdmult_modifier;
 
  double rb;
 
  int mb_energy;
  int sb_energy_level;
 
  ModeCosts mode_costs;
 
  MvCosts *mv_costs;
 
  IntraBCMVCosts *dv_costs;
 
  CoeffCosts coeff_costs;
 
  int errorperbit;
  int sadperbit;
 
  int seg_skip_block;
 
  int actual_num_seg1_blocks;
 
  int actual_num_seg2_blocks;
 
  int cnt_zeromv;
 
  int force_zeromv_skip;
 
  int prev_segment_id;
 
  // TODO(chiyotsai@google.com): Refactor this out of macroblock
  SuperBlockEnc sb_enc;
 
  CONTENT_STATE_SB content_state_sb;
 
  int pred_mv_sad[REF_FRAMES];
  int best_pred_mv_sad[2];
  int pred_mv0_sad[REF_FRAMES];
  int pred_mv1_sad[REF_FRAMES];
 
  uint8_t tpl_keep_ref_frame[REF_FRAMES];
 
  WARP_SAMPLE_INFO warp_sample_info[REF_FRAMES];
 
  int picked_ref_frames_mask[MAX_MIB_SIZE * MAX_MIB_SIZE];
 
  int nonrd_prune_ref_frame_search;
 
  PartitionSearchInfo part_search_info;
 
  int must_find_valid_partition;
 
  int skip_mode;
 
  int thresh_freq_fact[BLOCK_SIZES_ALL][MAX_MODES];
 
  WinnerModeStats *winner_mode_stats;
  int winner_mode_count;
 
  TXFM_RD_MODEL rd_model;
 
  // TODO(any): try to consolidate this speed feature with winner mode
  // processing.
  struct inter_modes_info *inter_modes_info;
 
  uint8_t compound_idx;
 
  COMP_RD_STATS comp_rd_stats[MAX_COMP_RD_STATS];
  int comp_rd_stats_idx;
 
  int recalc_luma_mc_data;
 
  IntraBCHashInfo intrabc_hash_info;
 
  int use_mb_mode_cache;
  const MB_MODE_INFO *mb_mode_cache;
  PixelLevelGradientInfo *pixel_gradient_info;
  bool is_sb_gradient_cached[PLANE_TYPES];
 
  unsigned int max_mv_context[REF_FRAMES];
 
  FullMvLimits mv_limits;
 
  TxfmSearchParams txfm_search_params;
 
  TxfmSearchInfo txfm_search_info;
 
  uint8_t color_sensitivity_sb[2];
  uint8_t color_sensitivity[2];
 
  unsigned int source_variance;
  unsigned int pred_sse[REF_FRAMES];
#if CONFIG_RT_ML_PARTITIONING
  DECLARE_ALIGNED(16, uint8_t, est_pred[128 * 128]);
#endif
 
  int try_merge_partition;
 
  Block4x4VarInfo *src_var_info_of_4x4_sub_blocks;
#ifndef NDEBUG
 
  SetOffsetsLoc last_set_offsets_loc;
#endif  // NDEBUG
} MACROBLOCK;
#undef SINGLE_REF_MODES
 
// Zeroes out 'n_stats' elements in the array x->winner_mode_stats.
// It only zeroes out what is necessary in 'color_index_map' (just the block
// size, not the whole array).
static INLINE void zero_winner_mode_stats(BLOCK_SIZE bsize, int n_stats,
                                          WinnerModeStats *stats) {
  // When winner mode stats are not required, the memory allocation is avoided
  // for x->winner_mode_stats. The stats pointer will be NULL in such cases.
  if (stats == NULL) return;
 
  const int block_height = block_size_high[bsize];
  const int block_width = block_size_wide[bsize];
  for (int i = 0; i < n_stats; ++i) {
    WinnerModeStats *const stat = &stats[i];
    memset(&stat->mbmi, 0, sizeof(stat->mbmi));
    memset(&stat->rd_cost, 0, sizeof(stat->rd_cost));
    memset(&stat->rd, 0, sizeof(stat->rd));
    memset(&stat->rate_y, 0, sizeof(stat->rate_y));
    memset(&stat->rate_uv, 0, sizeof(stat->rate_uv));
    // Do not reset the whole array as it is CPU intensive.
    memset(&stat->color_index_map, 0,
           block_width * block_height * sizeof(stat->color_index_map[0]));
    memset(&stat->mode_index, 0, sizeof(stat->mode_index));
  }
}
 
static INLINE int is_rect_tx_allowed_bsize(BLOCK_SIZE bsize) {
  static const char LUT[BLOCK_SIZES_ALL] = {
    0,  // BLOCK_4X4
    1,  // BLOCK_4X8
    1,  // BLOCK_8X4
    0,  // BLOCK_8X8
    1,  // BLOCK_8X16
    1,  // BLOCK_16X8
    0,  // BLOCK_16X16
    1,  // BLOCK_16X32
    1,  // BLOCK_32X16
    0,  // BLOCK_32X32
    1,  // BLOCK_32X64
    1,  // BLOCK_64X32
    0,  // BLOCK_64X64
    0,  // BLOCK_64X128
    0,  // BLOCK_128X64
    0,  // BLOCK_128X128
    1,  // BLOCK_4X16
    1,  // BLOCK_16X4
    1,  // BLOCK_8X32
    1,  // BLOCK_32X8
    1,  // BLOCK_16X64
    1,  // BLOCK_64X16
  };
 
  return LUT[bsize];
}
 
static INLINE int is_rect_tx_allowed(const MACROBLOCKD *xd,
                                     const MB_MODE_INFO *mbmi) {
  return is_rect_tx_allowed_bsize(mbmi->bsize) &&
         !xd->lossless[mbmi->segment_id];
}
 
static INLINE int tx_size_to_depth(TX_SIZE tx_size, BLOCK_SIZE bsize) {
  TX_SIZE ctx_size = max_txsize_rect_lookup[bsize];
  int depth = 0;
  while (tx_size != ctx_size) {
    depth++;
    ctx_size = sub_tx_size_map[ctx_size];
    assert(depth <= MAX_TX_DEPTH);
  }
  return depth;
}
 
static INLINE void set_blk_skip(uint8_t txb_skip[], int plane, int blk_idx,
                                int skip) {
  if (skip)
    txb_skip[blk_idx] |= 1UL << plane;
  else
    txb_skip[blk_idx] &= ~(1UL << plane);
#ifndef NDEBUG
  // Set chroma planes to uninitialized states when luma is set to check if
  // it will be set later
  if (plane == 0) {
    txb_skip[blk_idx] |= 1UL << (1 + 4);
    txb_skip[blk_idx] |= 1UL << (2 + 4);
  }
 
  // Clear the initialization checking bit
  txb_skip[blk_idx] &= ~(1UL << (plane + 4));
#endif
}
 
static INLINE int is_blk_skip(uint8_t *txb_skip, int plane, int blk_idx) {
#ifndef NDEBUG
  // Check if this is initialized
  assert(!(txb_skip[blk_idx] & (1UL << (plane + 4))));
 
  // The magic number is 0x77, this is to test if there is garbage data
  assert((txb_skip[blk_idx] & 0x88) == 0);
#endif
  return (txb_skip[blk_idx] >> plane) & 1;
}
 
#ifdef __cplusplus
}  // extern "C"
#endif
 
#endif  // AOM_AV1_ENCODER_BLOCK_H_